Research map / Image generation / Image restoration
AnyBokeh: Physics-Guided Any-to-Any Bokeh Editing with Optical Fingerprint Transfer
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes AnyBokeh, a two-stage physics-guided framework that estimates signed source CoC and disparity, transfers a source-specific optical fingerprint to a target focus/aperture, and runs a FLUX-based dual-CoC editor for any-to-any bokeh editing, plus a synthetic UnrealBokeh dataset.
Problem
Post-capture bokeh editing from a single image with arbitrary source defocus and unknown camera metadata; existing pipelines force an all-in-focus intermediate and require per-image bokeh-level calibration.
Approach
Estimate source signed CoC and disparity, fit a linear CoC-disparity slope (optical fingerprint), scale it by the aperture ratio to get the target CoC, and condition a generative editor on both source and target CoC maps for relative blur synthesis without AIF reconstruction or calibration.
Key results
- On any-to-any editing, AnyBokeh reaches LPIPS 0.4010 / DISTS 0.1426 / FID 24.60 on EBB! and 0.3033 / 0.0944 / 33.67 on RealBokeh, beating deblur-then-render baselines (best baseline DISTS 0.2588 on EBB!) without test-time calibration.
- On defocus deblurring, AnyBokeh has best DISTS and FID on EBB! (0.1378, 38.54) and RealBokeh (0.1033, 44.10), but worse LPIPS than Restormer/DRBNet.
- On AIF-to-bokeh rendering, AnyBokeh achieves best LPIPS, DISTS, and FID on both datasets (e.g., EBB! LPIPS 0.3615, FID 36.08).
Surprising findings
- AnyBokeh obtains worse LPIPS than restoration baselines on defocus deblurring, which the authors attribute to LPIPS penalizing plausible high-frequency recovery.
- Single-stage depth-based variant achieves slightly better LPIPS than the two-stage design, though two-stage is better on DISTS and FID.
Limitations
- Extremely defocused inputs lose high-frequency evidence, limiting sharp reconstruction.
- Very large bokeh balls reduced to small sizes can leave residual or hollow artifacts; rare in training data.
- Tiled high-resolution inference increases runtime.
Experimental setup
Datasets: UnrealBokeh EBB! RealBokeh
Benchmarks: EBB! RealBokeh
Metrics: LPIPS DISTS FID
Baselines: BokehMe BokehDiff Restormer DRBNet
Builds on: FLUX.1-Fill Bokeh Diffusion Genfocus LoRA Depth-Pro
Closest papers by meaning
- BokehFlow: Depth-Free Controllable Bokeh Rendering via Flow Matching (AAAI 2026 · similarity 0.88)
- Dr. Bokeh: DiffeRentiable Occlusion-aware Bokeh Rendering (CVPR 2024 · similarity 0.87)
- Towards Photorealistic and Efficient Bokeh Rendering via Diffusion Framework (CVPR 2026 · similarity 0.87)
- Any-to-Bokeh: Arbitrary-Subject Video Refocusing with Video Diffusion Model (ICLR 2026 · similarity 0.86)
- DOF-GS: Adjustable Depth-of-Field 3D Gaussian Splatting for Post-Capture Refocusing, Defocus Rendering and Blur Removal (CVPR 2025 · similarity 0.82)
- PixPerfect: Seamless Latent Diffusion Local Editing with Discriminative Pixel-Space Refinement (NeurIPS 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.